greyovo/PicQuery

🔍 Search local images with natural language on Android, powered by OpenAI's CLIP model. / 在 Android 上用自然语言搜索本地图片 (基于 OpenAI 的 CLIP 模型)

What it solves

PicQuery allows users to search through their local photo libraries using natural language descriptions (in English or Chinese) or by using another image as a query, without needing to upload photos to a cloud service.

How it works

The app builds a local index of photos on the device. It utilizes Apple's MobileCLIP2-S0 / dfndr2b models to create embeddings for images and text. It offers two different inference runtime options for Android: ONNX Runtime and TFLite (LiteRT), allowing users to compare search performance and accuracy across different backends.

Who it’s for

Android users who want a private, on-device way to search their photo galleries using AI-powered semantic search.

Highlights

  • On-device processing: All indexing and searching happen locally for privacy and speed.
  • Multilingual support: Supports search queries in both English and Chinese.
  • Image-based search: Ability to use a photo as a query to find similar images.
  • Flexible runtimes: Supports both ONNX Runtime and TFLite/LiteRT for model inference.

Related

  • Project
    egdels/makeacopy

    An offline, privacy-focused Android document scanner that uses on-device ML and OCR to create searchable PDFs for self-hosted workflows.

  • Project
    qpuchen/nnUNet_att_position_correction

    A semi-supervised 3D tooth segmentation framework that enhances nnU-Net with axial attention and a positional correction module to improve accuracy using limited labeled data.

  • Project
    rgeirhos/Stylized-ImageNet

    A tool for creating Stylized-ImageNet, a version of ImageNet that distorts textures to encourage CNNs to prioritize object shape over texture for better robustness.

  • Project
    IvanDrokin/torch-conv-kan

    TorchConv‑KAN is a PyTorch library that implements convolutional layers where each kernel is a set of learnable univariate functions (KAN, Fast‑KAN, Cheby‑KAN, etc.). It ships with many layer variants, CNN‑style model families (ResKANet, DenseKANet, VGG‑KAN, U‑Net‑KAN), pre‑trained Imagenet‑1k checkpoints, and training scripts that use Accelerate, Hydra, WandB, and Ray‑Tune. The project is under active development and accompanies a research paper on Kolmogorov‑Arnold convolutions.